Computational prediction of frequent hitters in target-based and cell-based assays

Compounds interfering with high-throughput screening (HTS) assay technologies (also known as “badly behaving compounds”, “bad actors”, “nuisance compounds” or “PAINS”) pose a major challenge to early-stage drug discovery. Many of these problematic compounds are “frequent hitters”, and we have recently published a set of machine learning models (“Hit Dexter 2.0”) for flagging such compounds.

Here we present a new generation of machine learning models which are derived from a large, manually curated and annotated data set. For the first time, these models cover, in addition to target-based assays, also cell-based assays. Our experiments show that cell-based assays behave indeed differently from target-based assays, with respect to hit rates and frequent hitters, and that dedicated models are required to produce meaningful predictions. In addition to these extensions and refinements, we explored a variety of additional setups for modeling, including the combination of four machine learning classifiers (i.e. k-nearest neighbors (KNN), extra trees, random forest and multilayer perceptron) with four sets of descriptors (Morgan2 fingerprints, Morgan3 fingerprints, MACCS keys and 2D physicochemical property descriptors).

Testing on holdout data as well as data sets of “dark chemical matter” (i.e. compounds that have been extensively tested in biological assays but have never shown activity) and known bad actors show that the multilayer perceptron classifiers in combination with Morgan2 fingerprints outperform other setups in most cases. The best multilayer perceptron classifiers obtained Matthews correlation coefficients of up to 0.648 on holdout data. These models are available via a free web service.